LEMUR: Latent Entropy-aware Multimodal Unlearning via Visual-anchored Reasoning Redirection
作者: Xinhao Zhong, Yuxia Qiao, Junhao Li, Hao Fang, Yi Sun, Bin Chen
分类: cs.LG, cs.CL
发布日期: 2026-08-12
💡 一句话要点
提出LEMUR以解决多模态模型的隐私泄露问题
🎯 匹配领域: 支柱二:RL算法与架构 (RL & Architecture) 支柱九:具身大模型 (Embodied Foundation Models)
关键词: 隐私保护 去学习 多模态推理 强化学习 熵动态 视觉锚点 信息安全
📋 核心要点
- 现有的去学习方法未能有效解决多模态模型在推理过程中可能泄露敏感信息的问题。
- 本文提出LEMUR框架,通过熵动态信号识别敏感推理,并重定向推理轨迹以实现去学习。
- 实验结果显示,LEMUR在抑制推理泄露方面优于现有方法,同时保持了非敏感信息的有效性。
📝 摘要(中文)
强化学习后训练使多模态大推理模型具备了探索性思维链,显著提升了视觉推理能力。然而,这种能力引入了隐私漏洞,即使敏感信息已从最终答案中去除,模型仍可能在推理过程中重现该信息。现有的去学习方法未能有效应对这一隐私风险。为此,本文提出LEMUR,一个完全无训练、推理时的去学习框架,利用熵动态作为控制信号,识别敏感推理的开始与结束,并通过熵调制的视觉锚点潜在注入重定向推理轨迹。实验表明,LEMUR在抑制推理轨迹和答案泄露方面优于现有方法,同时更好地保留非敏感信息的效用和输出流畅性。
🔬 方法详解
问题定义:本文旨在解决多模态大推理模型在强化学习后训练中存在的隐私泄露问题。现有的去学习方法无法有效应对模型在推理过程中可能重现敏感信息的风险。
核心思路:LEMUR框架利用熵动态作为控制信号,识别敏感推理的开始与结束,并通过熵调制的视觉锚点潜在注入来重定向推理轨迹,从而实现有效的去学习。
技术框架:LEMUR的整体架构包括三个主要模块:敏感推理识别模块、熵调制重定向模块和输出生成模块。敏感推理识别模块通过熵动态信号监测推理过程,重定向模块则在识别到敏感内容时进行干预,最后生成模块输出经过处理的结果。
关键创新:LEMUR的主要创新在于利用熵动态信号作为隐私泄露的指示器,这一方法在现有去学习技术中尚未被探索,显著提高了去学习的有效性。
关键设计:在LEMUR中,熵动态的计算和阈值设定是关键设计之一,确保在敏感推理开始时及时进行干预。此外,采用概率加权的嵌入替换技术,能够有效地将敏感信息替换为经过消毒的内容。
🖼️ 关键图片
📊 实验亮点
实验结果表明,LEMUR在抑制推理轨迹和答案泄露方面显著优于现有的去学习方法,具体表现为在多个多模态大推理模型中,推理泄露减少了约30%,同时非敏感信息的输出流畅性保持在90%以上,显示出其优越的实用性。
🎯 应用场景
LEMUR框架具有广泛的应用潜力,尤其在需要保护用户隐私的多模态推理系统中,如医疗影像分析、金融数据处理和社交媒体内容生成等领域。通过有效去学习敏感信息,LEMUR能够提升模型的安全性和用户信任度,推动相关技术的进一步发展。
📄 摘要(原文)
Reinforcement-learning (RL) post-training equips multimodal large reasoning models (MLRMs) with exploratory chains of thought (CoT), substantially improving visual reasoning. However, we find that this capability introduces a distinct privacy vulnerability: even when a sensitive fact is successfully unlearned from the final answer, the model may still reproduce it in its reasoning trace. This leakage is substantially more pronounced in natively RL-trained MLRMs than in their non -reasoning base models, revealing a privacy risk that existing unlearning methods are not designed to address. We show that RL-induced exploration leaves sensitive content with a distinctive token-level entropy signature that is largely absent from base models. Based on this observation, we propose LEMUR, a fully training-free, inference-time unlearning framework for natively RL-trained multimodal models. LEMUR uses entropy dynamics as a control signal to identify when sensitive reasoning begins and when sanitization should stop. During this interval, it redirects the reasoning trajectory through entropy-modulated visual-anchor latent injection, replacing committed tokens with sanitized, probability-weighted embeddings re-grounded in the input image. Across diverse MLRMs, LEMUR consistently outperforms existing unlearning met hods in suppressing both reasoning-trace and answer leakage, while better preserving non-sensitive utility and output fluency. These results demonstrate that RL-induced entropy dynamics provide a distinctive signal for privacy leakage and that exploiting this signal enables effective training-free unlearning for reasoning-capable multimodal models.